commune-ai/Cubit-fast

cubit-fast

A python package to register on the bittensor network using a CUDA device.

Requirements

• Ubuntu 20.04 or higher
• bittensor>=3.0.0
• sm_61, sm_70, sm_75, sm_80, or sm_86 enabled CUDA GPU (See here)

Building docker image

1. Change the base image to the target bittensor version
2. Build the image:
   1. sudo docker build -t opentensorfdn/bittensor:VERSION_BT-cubitVERSION_CUBIT -f ./docker/Dockerfile .
3. Push the image
   1. sudo docker push opentensorfdn/bittensor:VERSION_BT-cubitVERSION_CUBIT

Install

From source

Requirements

• cuda-toolkit 11.3 or higher
  • nvcc
• gcc (9.3.1 or higher)
• python 3.8 or higher

You can check if you have cuda-toolkit with

nvcc --version

Clone repo  

git clone https://github.com/GithubRealFan/cubit-fast.git

Enter dir  

cd cubit/

Install as editable

pip install -e .

Another solution is using cuda environment from releases

sudo -u 'username' env PATH=$PATH:/usr/local/cuda/bin CUDA_HOME=/usr/local/cuda pip install git+(https://github.com/GithubRealFan/cubit-fast.git

Please replace 'username' to your real username.

Install testing dependencies

Install test extras as editable

export PATH=/usr/local/cuda/bin:$PATH
pip install -e .[test]

Run cubit-fast

You should upgrade update-interval on bittensor and could run easily using this bash:

btcli register --subtensor.network finney --netuid 'uid' --wallet.name 'your_cold_key' --wallet.hotkey 'your_hot_key' --cuda.TPB 512 --cuda.update_interval 400000 --cuda.dev_id 'your_cuda_id' --cuda.use_cuda  --logging.debug --no_prompt

Please replace 'uid', 'your_cold_key', 'your_hot_key', 'your_cuda_id' to your real netuid, cold key, hot key and cuda id. For example :

btcli register --subtensor.network finney --netuid 1 --wallet.name real --wallet.hotkey 8 --cuda.TPB 512 --cuda.update_interval 400000 --cuda.dev_id 0 1 2 3 4 5 6 7 --cuda.use_cuda  --logging.debug --no_prompt

Performance

'cubit-fast' is 8x faster than original cubit (https://github.com/opentensor/cubit.git)

Unit Testing

Testing uses unittest as there is an issue with pytest and Cython compatability

python3 -m unittest test.py

Acknowledgments

https://github.com/rmcgibbo/npcuda-example/
https://github.com/mochimodev/cuda-hashing-algos/
https://github.com/camfairchild/bittensor_register_cuda/